Predicting the presence of diabetes in patients through diagnostic measurements¶

Phase 1: Data Preparation & Visualisation¶

Group Name: Group 67¶

Name and Student ID: Ting Li s3912985¶

Table of Contents¶

  • Introduction
    • Dataset Source
    • Dataset Details
    • Dataset Features
    • Target Feature
  • Goals & Objectives
  • Data Cleaning & Preprocessing
  • Data Exploration & Visualisation
  • Literature Review
  • Summary & Conclusions
  • References

Introduction ¶

Dataset Source ¶

The Diabetes dataset used in this study was sourced from Kaggle (UCI Machine Learning, 2017). This dataset contains diagnostic measurements of female patients of Pima Indian Heritage and whether or not they are diagnosed with diabetes within 5 years of the examination.

Dataset Details ¶

This dataset is about the prediction of diabetes in female patients of Pima Indian Heritage, using diagnostic measurements. There are 9 variables and 768 observations in this dataset. The outcome variable is about whether the patient has diabetes and the other 8 variables contain diagnostic measurements including the number of pregnancies, glucose test results, blood pressure, skin thickness, insulin, body mass index(BMI), Diabetes Pedigree Function, age.

Dataset Retrieval¶

The diabetes dataset has been downloaded from Kaggle. We will now load the data and modules to be used in Phase 1 of this study.

In [1]:
import warnings
warnings.filterwarnings("ignore")
In [2]:
# Import modules
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
import plotly.express as px
%matplotlib inline 
%config InlineBackend.figure_format = 'retina'
plt.style.use("seaborn")
from scipy.stats import normaltest

pd.set_option('display.max_columns', None)
In [3]:
diabetes = pd.read_csv("diabetes.csv")
diabetes.sample(10, random_state=999)
Out[3]:
Pregnancies Glucose BloodPressure SkinThickness Insulin BMI DiabetesPedigreeFunction Age Outcome
441 2 83 66 23 50 32.2 0.497 22 0
57 0 100 88 60 110 46.8 0.962 31 0
68 1 95 66 13 38 19.6 0.334 25 0
95 6 144 72 27 228 33.9 0.255 40 0
411 1 112 72 30 176 34.4 0.528 25 0
192 7 159 66 0 0 30.4 0.383 36 1
352 3 61 82 28 0 34.4 0.243 46 0
400 4 95 64 0 0 32.0 0.161 31 1
677 0 93 60 0 0 35.3 0.263 25 0
376 0 98 82 15 84 25.2 0.299 22 0

Dataset Features ¶

The table below provides descriptions of features in the diabetes dataset, which are taken from Kaggle (UCI Machine Learning, 2017).

In [4]:
from tabulate import tabulate

table = [['Name','Data Type','Units','Description of Features'],
         ['Pregnancies','Numeric','NA','Number of times pregnant'],
         ['Glucose','Numeric','NA','Plasma glucose concentration a 2 hours in an oral glucose tolerance test (GTT)'],
         ['BloodPressure','Numeric','mm Hg','Diastolic blood pressure'],
         ['SkinThickness','Numeric','mm','Triceps skin fold thickness'],
         ['Insulin','Numeric','mu U/ml','2-Hour serum insulin'],
         ['BMI','Numeric','NA','Body mass index (weight in kg/(height in m)^2)'],
         ['DiabetesPedigreeFunction','Numeric','NA','Diabetes pedigree function (calculated on diabetes history in relatives and their genetic relationship)'],
         ['Age','Numeric','km','Distance from Central Business District'],
         ['Outcome','Binary','NA','whether a patient has diabetes within 5 years of the examination(class variable, 0 or 1)']]
table_name = "Table 1: Descriptions of Features"
print(table_name)
print(tabulate(table, headers='firstrow', tablefmt='grid'))

Table 1: Descriptions of Features
+--------------------------+-------------+---------+---------------------------------------------------------------------------------------------------------+
| Name                     | Data Type   | Units   | Description of Features                                                                                 |
+==========================+=============+=========+=========================================================================================================+
| Pregnancies              | Numeric     | NA      | Number of times pregnant                                                                                |
+--------------------------+-------------+---------+---------------------------------------------------------------------------------------------------------+
| Glucose                  | Numeric     | NA      | Plasma glucose concentration a 2 hours in an oral glucose tolerance test (GTT)                          |
+--------------------------+-------------+---------+---------------------------------------------------------------------------------------------------------+
| BloodPressure            | Numeric     | mm Hg   | Diastolic blood pressure                                                                                |
+--------------------------+-------------+---------+---------------------------------------------------------------------------------------------------------+
| SkinThickness            | Numeric     | mm      | Triceps skin fold thickness                                                                             |
+--------------------------+-------------+---------+---------------------------------------------------------------------------------------------------------+
| Insulin                  | Numeric     | mu U/ml | 2-Hour serum insulin                                                                                    |
+--------------------------+-------------+---------+---------------------------------------------------------------------------------------------------------+
| BMI                      | Numeric     | NA      | Body mass index (weight in kg/(height in m)^2)                                                          |
+--------------------------+-------------+---------+---------------------------------------------------------------------------------------------------------+
| DiabetesPedigreeFunction | Numeric     | NA      | Diabetes pedigree function (calculated on diabetes history in relatives and their genetic relationship) |
+--------------------------+-------------+---------+---------------------------------------------------------------------------------------------------------+
| Age                      | Numeric     | km      | Distance from Central Business District                                                                 |
+--------------------------+-------------+---------+---------------------------------------------------------------------------------------------------------+
| Outcome                  | Binary      | NA      | whether a patient has diabetes within 5 years of the examination(class variable, 0 or 1)                |
+--------------------------+-------------+---------+---------------------------------------------------------------------------------------------------------+

Target Features ¶

For this project, the target feature in this dataset will be the outcome of patients' diagnosis, that is, prediction of whether or not the patient will develop diabetes in 5 years based on 8 descriptive variables.

Goals & Objectives ¶

Diabetes is a major cause of diseases and conditions such as heart attacks, stroke, kidney failure and lower limb amputation (World Health Organization, 2023). Prevalence of diabetes can vary by region or ethnic groups and the population of Pima Indian Heritage (Schulz et al., 2006). A model that can predict diabetes using diagnostic measurement of subjects can be helpful in real-world medical situations. For example, it can be used to assist detection of diabetes from individuals' health check-up data, or to be used to predict other diseases and conditions for early prevention and intervention.

The main objective of this project is to predict whether diabetes will be onset in a patient in 5 years, based on measurements of 8 risk factors for diabetes in an examination. For Phase 1 of the project, we are going to: (1) perform data cleaning and preprocessing, (2) explore patterns and relationships within the dataset through descriptive statistics and data visualisation, and (3) provide literature review on the use of machine learning methods in the prediction and diagnosis of diseases.

Data selected for this dataset satisfies the following criteria: (1) all subjects are females of Pima Indian Heritage, aged 21 or above during time of the examination; (2) for each subject only one examination was selected and the chosen examination gave a non-diabetic glucose tolerance test measurement, that is below 200; (3) subjects that are diagnosed with diabetes within 1 year of examination are excluded from the dataset to remove easy-to-detect data (Smith, Everhart, Dickson, Knowler, & Johannes, 1988).

Data Cleaning & Preprocessing ¶

This section includes the data cleaning and preprocessing steps undertaken to prepare the data for predictive modeling in Phase 2 of this project. These steps are listed below and will be performed in order as specified in the course website (www.featureranking.com).

Data Cleaning Steps

  • Dataset observation
    • Check and drop irrelevant features in our dataset
    • Check and rename/ modify some column names if needed
  • Outliers, unusual values and missing values
    • Check for outliers and unusual values
    • Impute or drop all rows with outliers and unusual values
    • Check for missing values
    • Impute or drop all rows with missing values
  • Check if the target feature is label-encoded

One-hot-encoding (OHE) should be performed for categorical descriptive feature to encode them to be numeric, therefore is not a part of the project in Phase 1 & 2 as there is no categorical data in this dataset. Scaling of numeric features is not included in Phase 1 as specified in Phase 1 criteria.

Dataset Observation¶

Column names of the diabetes dataset are displayed below.

In [5]:
diabetes.columns
Out[5]:
Index(['Pregnancies', 'Glucose', 'BloodPressure', 'SkinThickness', 'Insulin',
       'BMI', 'DiabetesPedigreeFunction', 'Age', 'Outcome'],
      dtype='object')

Upon inspection of features, no redundancy was found. However, column BloodPressure, SkinThickness, and DiabetesPedigreeFunction are not named properly, and should be renamed with underscores for readability. After renaming, we observe from the sample that changes have been made and all column names are consistent.

In [6]:
#mapping new column names 

columns_new = {
    'BloodPressure': 'Blood_Pressure',
    'SkinThickness': 'Skin_Thickness',
    'DiabetesPedigreeFunction': 'Diabetes_Pedigree_Function'}

# rename columns
diabetes = diabetes.rename(columns = columns_new)
diabetes.sample(5, random_state=999)
Out[6]:
Pregnancies Glucose Blood_Pressure Skin_Thickness Insulin BMI Diabetes_Pedigree_Function Age Outcome
441 2 83 66 23 50 32.2 0.497 22 0
57 0 100 88 60 110 46.8 0.962 31 0
68 1 95 66 13 38 19.6 0.334 25 0
95 6 144 72 27 228 33.9 0.255 40 0
411 1 112 72 30 176 34.4 0.528 25 0

Then we continue to check the dataset for its shape and data types to make sure they are suitable and match the description.

In [7]:
print(f"Shape of the dataset = {diabetes.shape} \n")
print(f"Data types are below where 'object' indicates a string type: ")
print(diabetes.dtypes)

Shape of the dataset = (768, 9) 

Data types are below where 'object' indicates a string type: 
Pregnancies                     int64
Glucose                         int64
Blood_Pressure                  int64
Skin_Thickness                  int64
Insulin                         int64
BMI                           float64
Diabetes_Pedigree_Function    float64
Age                             int64
Outcome                         int64
dtype: object

We observe that the shape of the dataset matches the description and data types are suitable. Since all variables are numeric, we now move on to generating summary statistics for all variables.

Outliers, Unusual Values and Missing Values¶

Table 2 shows a summary of numerical features all features in the diabetes dataset. From the summary table we found unusual values in variables below:

  • Glucose
  • Blood_Pressure
  • Skin_Thickness
  • Insulin
  • BMI

As subjects of the diabetes dataset are living human beings, all above variables can not take 0 as a value. We observed below that there are no missing values in the diabetes dataset. It is likely that 0 was used during data entry to represent missing values.

In [8]:
from IPython.display import display, HTML
display(HTML('<b>Table 2: Summary of numerical features</b>'))
diabetes.describe(include=['int64','float64']).T
Table 2: Summary of numerical features
Out[8]:
count mean std min 25% 50% 75% max
Pregnancies 768.0 3.845052 3.369578 0.000 1.00000 3.0000 6.00000 17.00
Glucose 768.0 120.894531 31.972618 0.000 99.00000 117.0000 140.25000 199.00
Blood_Pressure 768.0 69.105469 19.355807 0.000 62.00000 72.0000 80.00000 122.00
Skin_Thickness 768.0 20.536458 15.952218 0.000 0.00000 23.0000 32.00000 99.00
Insulin 768.0 79.799479 115.244002 0.000 0.00000 30.5000 127.25000 846.00
BMI 768.0 31.992578 7.884160 0.000 27.30000 32.0000 36.60000 67.10
Diabetes_Pedigree_Function 768.0 0.471876 0.331329 0.078 0.24375 0.3725 0.62625 2.42
Age 768.0 33.240885 11.760232 21.000 24.00000 29.0000 41.00000 81.00
Outcome 768.0 0.348958 0.476951 0.000 0.00000 0.0000 1.00000 1.00
In [9]:
# Count missing values in each column
print(f"\nNumber of missing values for each column/ feature:")
print(diabetes.isnull().sum())

Number of missing values for each column/ feature:
Pregnancies                   0
Glucose                       0
Blood_Pressure                0
Skin_Thickness                0
Insulin                       0
BMI                           0
Diabetes_Pedigree_Function    0
Age                           0
Outcome                       0
dtype: int64

To determine the method used to deal with missing values in the dataset, we checked the count of zeros in column Glucose, Blood_Pressure, Skin_Thickness, Insulin, and BMI. As shown above, the count of missing values in those columns range from 5 to 374. As the dataset only has 768 observations, dropping all columns with missing values will result in a loss of valuable information. Therefore, missing values should be imputed for this project.

In [10]:
# Create subset for inspection of 5 columns with zeros
subset_cols_with_0 = diabetes.iloc[:,1:6]
count_0 = subset_cols_with_0.eq(0).sum()
count_0
Out[10]:
Glucose             5
Blood_Pressure     35
Skin_Thickness    227
Insulin           374
BMI                11
dtype: int64

Numeric values can be imputed with the mean or median (Taheri, 2021). The median is often used for imputation when data is skewed and the mean is used when data is normally distrubuted, as the mean is more sensitive to outliers than the median. We will inspect the distribution of Glucose, Blood_Pressure, Skin_Thickness, Insulin, and BMI to decide which imputation to use.

In [11]:
p1 = subset_cols_with_0.hist(figsize = (20,20))
plt.suptitle('Figure 1: Histogram of Glucose, Blood_Pressure, Skin_Thickness, Insulin, and BMI', fontsize=20)
Out[11]:
Text(0.5, 0.98, 'Figure 1: Histogram of Glucose, Blood_Pressure, Skin_Thickness, Insulin, and BMI')

Histogram (Figure 1) above showed skewness in all 5 variables. For certainty, we will perform normality test below.

In [12]:
# Normality test
norm = subset_cols_with_0.apply(lambda x: normaltest(x))
print(norm)

     Glucose  Blood_Pressure  Skin_Thickness       Insulin           BMI
0  12.385057    3.058869e+02       17.344871  3.875778e+02  8.614248e+01
1   0.002045    3.780127e-67        0.000171  6.895343e-85  1.969687e-19

The second row of normality test result shows p-value for each variable. As all 5 p-values are smaller than the default significance level (alpha = 0.05), we reject the null hypothesis and conclude that data is not normally distributed for all 5 columns.

Therefore, all missing values will be imputed with the median for the diabetes dataset.

In [13]:
# Impute with median
diabetes.loc[diabetes['Glucose'] == 0, 'Glucose'] = diabetes['Glucose'].median()
diabetes.loc[diabetes['Blood_Pressure'] == 0, 'Blood_Pressure'] = diabetes['Blood_Pressure'].median()
diabetes.loc[diabetes['Skin_Thickness'] == 0, 'Skin_Thickness'] = diabetes['Skin_Thickness'].median()
diabetes.loc[diabetes['Insulin'] == 0, 'Insulin'] = diabetes['Insulin'].median()
diabetes.loc[diabetes['BMI'] == 0, 'BMI'] = diabetes['BMI'].median()

# Check data after imputation
display(HTML('<b>Table 3: Summary of numerical features after imputation</b>'))
diabetes.describe(include=['int64','float64']).T
Table 3: Summary of numerical features after imputation
Out[13]:
count mean std min 25% 50% 75% max
Pregnancies 768.0 3.845052 3.369578 0.000 1.00000 3.0000 6.00000 17.00
Glucose 768.0 121.656250 30.438286 44.000 99.75000 117.0000 140.25000 199.00
Blood_Pressure 768.0 72.386719 12.096642 24.000 64.00000 72.0000 80.00000 122.00
Skin_Thickness 768.0 27.334635 9.229014 7.000 23.00000 23.0000 32.00000 99.00
Insulin 768.0 94.652344 105.547598 14.000 30.50000 31.2500 127.25000 846.00
BMI 768.0 32.450911 6.875366 18.200 27.50000 32.0000 36.60000 67.10
Diabetes_Pedigree_Function 768.0 0.471876 0.331329 0.078 0.24375 0.3725 0.62625 2.42
Age 768.0 33.240885 11.760232 21.000 24.00000 29.0000 41.00000 81.00
Outcome 768.0 0.348958 0.476951 0.000 0.00000 0.0000 1.00000 1.00

Table 3 indicates that imputation has been performed successfully, and no unusual value was found for all variables except Pregnancies. We should notice that the maximum number of pregnancies is 17, and inspect the age of the that subject.

Also, for the context of this dataset, it is worth noticing that the maximum value for Glucose is 199, which fits the criteria that glucose test measurement for all subjects should be non-diabetic, which is below 200.

In [14]:
# Find the age of the subject who has been pregnant 17 times
value_age = diabetes.loc[diabetes['Pregnancies'] == 17, 'Age'].values[0]
print(f"The subject who has been pregnant 17 times is {value_age} years old.\n")
# Check if there is pregnancy at an unusual age
subset_unusual_age_pregnancy = diabetes[(diabetes['Pregnancies'] - diabetes['Age']).abs() < 15][['Pregnancies', 'Age']]
subset_unusual_age_pregnancy

The subject who has been pregnant 17 times is 47 years old.
Out[14]:
Pregnancies Age
731 8 22

As seen above, it is not abnormal that a 47 years old female can be pregnant 17 times. We also checked if there is anyone pregnant at an unusual age, for example, a 21 years old subject with 15 times pregnancy is likely to be a data entry error. Upon inspection of the dataset, where the difference of subjects' age and times of pregnancies is smaller than 15, we found patient 731 with 8 pregnancies at age 22. As it is possible for a female to be fertile at 14 years old, we conclude that there is no unusual values in the dataset.

Check Target Feature¶

In [15]:
# Count each unique value within the target feature
outcome_counts = diabetes['Outcome'].value_counts()
# Calculate the percentage of each unique value within the target feature
outcome_percentage = outcome_counts/outcome_counts.sum()*100
outcome_percentage
print(outcome_counts)
print(outcome_percentage)

0    500
1    268
Name: Outcome, dtype: int64
0    65.104167
1    34.895833
Name: Outcome, dtype: float64

As specified in Table 1, the target feature Outcome has been label-encoded to 0 and 1. Out of all 768 patients, 268 was diagnosed with diabetes within 5 years of the examination chosen for the dataset, which is 34.9% of all patients. Diabetes was not onset for the rest of the dataset, which is 500 patients and 65.1% of all patients.

Data Exploration & Visualisation ¶

After all data cleaning steps, we believe that the diabetes dataset is ready to be used for visualisation and exploration.

Univariate Visualisation¶

First we inspect the distribution of all variables in the dataset from the histogram below. The shape of distribution fits descriptive statistics in Table 3. We also notice in Skin_Thickness and Insulin where the missing data was imputed with the median in previous data cleaning steps, shown on the lower end of the distribution. We now move on to inspect variable BMI and Insulin in more detail.

In [16]:
p2 = diabetes.hist(figsize = (20,20))
plt.suptitle('Figure 2: Histogram of all variables', fontsize=20)
Out[16]:
Text(0.5, 0.98, 'Figure 2: Histogram of all variables')

Figures 3A and 3B shows the distribution of BMI where the majority of data falls above 25 and more than half above 30, indicating a body mass index of overweight (25) or obesity (30) in subjects of the diabetes dataset as suggested by Centres for Disease Control and Prevention (CDC).

In [17]:
# Boxplot of BMI
plt.figure(figsize = (15,8))
sns.boxplot(diabetes['BMI']).set_title('Figure 3A: Box Plot of BMI', fontsize = 15)
plt.show();
In [18]:
# Histogram of BMI
plt.figure(figsize = (15,8))
sns.distplot(diabetes['BMI'], kde=True, bins=40).set_title('Figure 3B: Histogram of BMI', fontsize = 15)
plt.show();

Figures 4A and 4B shows the distribution of Insulin. Data has a wide range and shows right-skewness. Therefore, data transformation such as log transformation might be needed in Phase 2, for better performance of machine learning algorithms that are more sensitive to normality.

In [19]:
# Boxplot of Insulin
plt.figure(figsize = (15,8))
sns.boxplot(diabetes['Insulin']).set_title('Figure 4A: Box Plot of Insulin', fontsize = 15)
plt.show();
In [20]:
# Histogram of Insulin
plt.figure(figsize = (15,8))
sns.distplot(diabetes['Insulin'], kde=True, bins=40).set_title('Figure 4B: Histogram of Insulin', fontsize = 15)
plt.show();

Two-variable Visualisation¶

We now move on to investigate relationship between each two variables. Figure 5 shows a scatter matrix for all variables in the diabetes data. Visual inspection of Figure 5 indicates a weak to moderate linear relationship between BMI and Skin_Thickness, and Pregnancies and Age.

In [21]:
sns.pairplot(diabetes)
plt.suptitle('Figure 5: Scatter matrix of diabetes data', y = 1.05, fontsize=25)
Out[21]:
Text(0.5, 1.05, 'Figure 5: Scatter matrix of diabetes data')

To further investigate the correlation between between each variables and measure strength of linear relationship, a heatmap was plotted in Figure 6 of correlation coefficients coded by color. With the cmap parameter set to 'Blues', a correlation coefficient r=1 is represented with dark blue, and the blue shade gets lighter as the correlation coefficient decreases, indicating a weaker linear relationship.

Figure 6 shows moderate positive linear relationship between BMI and Skin_Thickness (r=0.55), and between Pregnancies and Age (r=0.54) which supports our inspection of Figure 5. Other pairs that have moderate positive linear relationship include Glucose and Insulin (r=0.36), Blood_Pressure and Age (r=0.32), Glucose and Outcome (r=0.49), and BMI and Outcome (r=0.31). Weak linear relationship exists for the rest of paired variables with both positive and negative correlations.

Since body mass index (BMI), Age, family history (measured by Diabetes_Pedigree_Function), and high blood pressure (Blood_Pressure) are found to be high risk factors for diabetes (CDC, n.d.), we will visualise these variables and explore their relationship with the Glucose(GTT measurement for diabetes).

In [22]:
corr_matrix = diabetes.corr()
sns.heatmap(corr_matrix, cmap = 'Blues', annot = True).set_title('Figure 6: Heatmap of diabetes data', fontsize = 15)
plt.show()

Figure 7 shows the relationship between BMI and Glucose, and Figure 8 shows the relationship between Diabetes_Pedigree_Function and Glucose. Both Figure 7 and 8 indicates a weak positive linear relationship. However, it is unclear how the output differs for each class of the target feature Outcome.

We will now include the Outcome variable in the visualisation and produce three-variable plots.

In [23]:
# BMI and Glucose
plt.figure(figsize = (15,8))
plt.scatter(diabetes['BMI'], diabetes['Glucose'], alpha = 0.3)
plt.title('Figure 7: Scatterplot of Glucose and BMI', fontsize = 15)
plt.xlabel('BMI')
plt.ylabel('Glucose')
plt.show();
In [24]:
# Diabetes_Pedigree_Function and Glucose
plt.figure(figsize = (15,8))
plt.scatter(diabetes['Diabetes_Pedigree_Function'], diabetes['Glucose'], alpha = 0.3)
plt.title('Figure 8: Scatterplot of Glucose and Diabetes_Pedigree_Function', fontsize = 15)
plt.xlabel('Diabetes_Pedigree_Function')
plt.ylabel('Glucose')
plt.show();

Three-variable Visualisation¶

As stated in the previous section, Outcome is added to three-variable visualisations, where different classes within Outcome are plotted as different colors. Patients who are diagnosed with diabetes within 5 years of examination have Outcome variable equals 1, which are colored yellow. Patients who were not found with diabetes after 5 years of examination are colored blue, with a value 0 for the Outcome variable.

Figure 9 shows a three-variable scatter plot for Glucose, BMI and Outcome. The output for both Glucose and BMI is relatively higher for patients diagnosed with diabetes within 5 years, where Outcome=1.

In [25]:
# BMI, Glucose, and Outcome
import plotly.io as pio
pio.renderers.default = "notebook_connected"
p3 = px.scatter(diabetes, x="BMI", y="Glucose",
                color="Outcome",
                  log_x=True, size_max=60)
p3.update_layout(
    title="Figure 9: Scatterplot of Glucose vs BMI by Outcome",
    xaxis_title="BMI",
    yaxis_title="Glucose"
)

p3.show()

Figure 10 shows a three-variable scatter plot for Diabetes_Pedigree_Function, Glucose and Outcome. The output for Glucose is relatively higher for patients colored in yellow, whose Outcome variable equals 1, but the difference is not as evident for the Diabetes_Pedigree_Function variable.

In [26]:
# Diabetes_Pedigree_Function, Glucose and Outcome
p3_5 = px.scatter(diabetes, x="Diabetes_Pedigree_Function", y="Glucose",
	   
                color="Outcome",
                  log_x=True, size_max=60)
p3_5.update_layout(
    title="Figure 10: Scatterplot of Glucose vs Diabetes_Pedigree_Function by Outcome",
    xaxis_title="Diabetes_Pedigree_Function",
    yaxis_title="Glucose"
)

p3_5.show()

Figure 11 is a 3D scatter plot of Blood_Pressure, Age, Glucose by Outcome. It shows a weak tendency where subjects in yellow tend to cluster towards the higher values of Blood_Pressure, Age, and Glucose, suggesting a weak positive linear relationship of Blood_Pressure, Age, and Glucose with the outcome of diabetes prediction in the dataset.

In [27]:
plt.figure(figsize = (15,8))
p4=px.scatter_3d(diabetes, x ='Age', y ='Blood_Pressure', z='Glucose', color = 'Outcome')
p4.update_layout(title="Figure 11: 3D scatterplot of Glucose vs Blood_Pressure vs Age, by Outcome")
p4.show()

<Figure size 1500x800 with 0 Axes>

Literature Review ¶

Introduction¶

The prevalence of diabetes has been rising rapidly in the past few decades, affecting people from all over the world, especially those in low and middle-income countries. Machine learning algorithms can be used to predict risks of having diabetes and other aspects of diabetes management.

Predicting Risks of Developing Diabetes¶

There are many risk factors associated with the development of diabetes, and machine learning have been using to predict the risk of developing diabetes in individuals over the years.

For example, support vector machine(SVM) (Abbas et al, 2019), decision tree (Cahn et al, 2020), logistic regression (Choi et al., 2019), K-nearest neighbour (KNN) (Farran et al, 2019), forward neural network (Nguyen et al., 2019) and farthest first (Farhana, 2021) have been used to predict the risk of getting diabetes based on risk factors such as obesity, high blood pressure, and kidney disease and measurements such as oral glucose tolerance test results.

Surprisingly, linear regression models resulted in a better performance than 8 other machine learning algorithms in a study by Ye et al. (2020), for the prediction of gestational diabetes results, followed by gradient boosting decision trees.

Pima Indian Diabetes Dataset¶

The Pima Indian diabetes dataset has been used since the last century to study the application of machine learning algorithm in predicting diabetes (Smith et al. 1988). In 2022, Dahiya et al. found random forest to be the most accurate model for the Pima Indian diabetes dataset, with a precision of 85.06%, compared to other machine learning algorithms. Sarathambekai et al. (2022) reached a similar conclusion when comparing random forest to decision tree, logistic regression, SVM, KNN, G-Naïve Bayes and gradient boost.

Other Uses of Machine Learning in Diabetes Management¶

Machine learning models can be applied in the management and treatment of diabetes. Singla et al. (2022) used random forest algorithms to create decision trees, to support clinical decision making when prescribing diabetes medications. The model used data including demographics of patients, previous medications, biochemical markers and clinical parameters for the patient, and can be used to develop standardised diabetes care and potentially improve patients’ clinical outcome.

Maritsch et al. (2020) developed a machine learning model using gradient boosting decision trees to detect hypoglycaemia according to smartwatch sensor data. It can be sued to monitor glucose level in blood to prevent hypoglycaemia as a diabetic complication.

Another use of machine learning is to assist in predicting high risk individuals, for better allocation or medical resources and targeted preventions, as diabetes and complications of diabetes such as heart and kidney disease create a burden on the healthcare system and patients’ quality of life. Aminian et al. (2020) used random forest to develop individualised risk scores for each diabetic complications, with or without metabolic surgery, which is a preventative measure.

Deep learning algorithms can be used for early diagnosis and treatment to improve outcomes of diabetes complications. A model based on convolutional neural network (CNN) was developed by Gulshan et al. (2016) to recognise diabetic retinopathy, which is a complication that can progress to blindness. The model has up to 97.5% sensitivity for analysis of fundoscopic images. The model would help General Practitioners (GPs) in considering subsequent treatments or even develop a standardised screening for diabetic retinopathy.

CNN was also used for forecasting kidney function using ultrasound imaging of kidney, with RestNet models to help manage diabetes related kidney conditions (Kuo et al., 2019). It was found to be the most cost-efficient model compared to common machine learning approaches such as histogram of oriented gradients and local binary pattern. Result from the model was then inputted to XGBoost models for fine-tuning.

Conclusion¶

The application of machine learning provides a new tool in assisting medical practitioners in many aspects of medicine, including risk prediction, diagnosis and treatment selections. As discussed above, it has a potential to revolutionise prevention and care of diabetes.

Summary & Conclusions ¶

Phase 1 of the project covered data cleaning, data preprocessing, data visualisation and exploration. After initial data observation, 3 Variables were renamed for readability. Then, the dataset was checked for outliers, unusual values and missing values, and they were found for variable Glucose, Blood_Pressure, Skin_Thickness, Insulin, and BMI. Median imputation was used as all 5 variables have skewed distribution. Upon further inspection on all variables the dataset was deemed clean and ready for visualisation and exploration.

As some of the data are skewed, we may need to perform scaling and transformation in Phase 2 of this project for better performance of machine learning algorithms that are more sensitive to non-normality. We have seen moderate correlations between BMI and Skin_Thickness and between Pregnancies and Age in the dataset. Glucose and BMI are found to be moderately correlated with the outcome of diabetes prediction in the dataset.

More visualisations were generated to explore the relationship between target feature Outcome with BMI, Age, family history (measured by Diabetes_Pedigree_Function), and high blood pressure (Blood_Pressure). We found changes in output when comparing relationships of descriptive variables between two classes of the Outcome variable, since patients who are diagnosed with diabetes within 5 years (Outcome = 1) tend to take a higher value than those who were not diabetic in 5 years, especially for their value of Glucose.

Further analysis and modeling will be done in Phase 2 to develop a reliable model for forecasting diabetes with the dataset.

References ¶

Abbas, H.T., Alic, L., Erraguntla, M., Ji, J.X., Abdul-Ghani, M., Abbasi, Q.H., Qaraqe, M.K. (2019). Predicting long-term type 2 diabetes with support vector machine using oral glucose tolerance test. PLoS One, 14(e0219636). https://doi.org/10.1371/journal.pone.0219636

Akman, D. (2022). Phase1_Report_Sample.ipynb [Jupyter Notebook]. GitHub. https://github.com/akmand/stats_tutorials/blob/main/Phase1_Report_Sample.ipynb

Aminian, A., Zajichek, A., Arterburn, D.E., Wolski, K.E., Brethauer, S.A., Schauer, P.R., Nissen, S.E., Kattan, M.W. (2020). Predicting 10-Year Risk of End-Organ Complications of Type 2 Diabetes With and Without Metabolic Surgery: A Machine Learning Approach. Diabetes Care, 43 (4), 852–859. https://doi.org/10.2337/dc19-2057

Cahn, A., Shoshan, A., Sagiv, T., Yesharim, R., Goshen, R., Shalev, V., Raz, I. (2020). Prediction of progression from pre-diabetes to diabetes: development and validation of a machine learning model. Diabetes Metab Res Rev, 36(2). doi: 10.1002/dmrr.3252

Centers for Disease Control and Prevention. (n.d.). Search results for diabetes. https://search.cdc.gov/search/?query=diabetes&dpage=1

Choi, B.G., Rha, S.W., Kim, S.W., Kang, J.H., Park, J.Y., Noh, Y.K. (2019). Machine learning for the prediction of new-onset diabetes mellitus during 5-year follow-up in non-diabetic patients with cardiovascular risks. Yonsei Med J, 60(2), 191–199. doi: 10.3349/ymj.2019.60.2.191.

Dahiya, N., Gupta, S., Garg, M. (2022). Diabetes screening using machine learning algorithm. AIP Conference Proceedings, 2357(100019). https://aip-scitation-org.ezproxy.lib.rmit.edu.au/doi/abs/10.1063/5.0080850

Farhana, B., Munidhanalakshmi, K., MadanaMohana, R.(2021). Predict Diabetes Mellitus Using Machine Learning Algorithms. Journal of Physics: Conference Series. 2089, 012002. doi:10.1088/1742-6596/2089/1/012002

Farran, B., AlWotayan, R., Alkandari, H., Al-Abdulrazzaq, D., Channanath,A., Thanaraj, T.A. (2019). Use of noninvasive parameters and machine-learning algorithms for predicting future risk of type 2 diabetes: a retrospective cohort study of health data from Kuwait. Front Endocrinol, 10(624). doi: 10.3389/fendo.2019.00624

Georgiou, S. (2023). Topic1_slides.pdf [PDF file]. Canvas@RMIT University. https://rmit.instructure.com/courses/112586/files/29073439?wrap=1

Gulshan, V., Peng, L., Coram, M., Stumpe M.C., Wu, D., Narayanaswamy, A., Venugopalan. S., Widner, K., Madams, T., Cuadros, J., Kim, R., Raman, R., Nelson, P.C., Mega, J.L., Webster, D.R. (2016) Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs, JAMA, 316(22),2402-2410. doi:10.1001/jama.2016.17216

Kuo, C., Chang, C., Liu, K., Lin, W., Chiang, H., Chung, C., Ho, M., Sun, P., Yang, R., Chen, K. (2019). Automation of the kidney function prediction and classification through ultrasound-based kidney imaging using deep learning. Npj Digital Medicine, 29(2), https://www.nature.com/articles/s41746-019-0104-2

Maritsch, M., Föll, S., Lehmann, V., Bérubé, C., Kraus, M., Feuerriegel, S., Kowatsch, T., Züger, T., Stettler, C., Fleisch, E., Wortmann, F. (2020). Towards Wearable-based Hypoglycemia Detection and Warning in Diabetes. Extended Abstracts of the 2020 CHI Conference on human factors in computing systems (pp. 1-8). Association for Computing Machinery. http://dx.doi.org/10.1145/3334480.3382808

Nguyen, B.P., Pham, H.N., Tran, H.,Nghiem, N., Nguyen, Q.H., Do, T.T.T, Tran, C.T., Simpson, C.R. (2019). Predicting the onset of type 2 diabetes using wide and deep learning with electronic health records. Comput Methods Programs Biomed, 182(105055). doi: 10.1016/j.cmpb.2019.105055..

Sarathambekai, S., Vairam, T., Sathya Seelan, K., Prem Kumar, R., Rakshith, Ram Varun, Jawhar, M. G. (2022). A Comparative Analysis of Diabetes Prediction Models using Machine Learning Algorithms. 8th International Conference on Advanced Computing and Communication Systems (ICACCS), (pp. 261-265). doi: 10.1109/ICACCS54159.2022.9785280.

Singla, R., Aggarwal, S., Bindra, J., Garg, A. (2022) Developing Clinical Decision Support System using Machine Learning Methods for Type 2 Diabetes Drug Management. Indian Journal of Endocrinology and Metabolism, 26(1),(pp. 44-49). DOI: 10.4103/ijem.ijem_435_21

Smith, J. W., Everhart, J. E., Dickson, W. C., Knowler, W. C., & Johannes, R. S. (1988). Using the ADAP learning algorithm to forecast the onset of diabetes mellitus. In Proceedings of the Symposium on Computer Applications and Medical Care. IEEE Computer Society Press, 261-265. https://www.astro.uvic.ca/~karun/UVic_ACML/Smith1988.pdf

Taheri, S (2021, Oct 1) Module 5 Notes, Data Wrangling [Course website]. http://rare-phoenix-161610.appspot.com/secured/Module_05.html.

Talaei-Khoei, A., Wilson, J.M., (2018). Identifying people at risk of developing type 2 diabetes: a comparison of predictive analytics techniques and predictor variables. Int J Med Inform, 119, 22–38. DOI: 10.1016/j.ijmedinf.2018.08.008

World Health Organization. (2021, November 4). Diabetes. https://www.who.int/news-room/fact-sheets/detail/diabetes

Ye, Y., Xiong, Y., Zhou, Q., Wu, J., Li, X., Xiao, X., (2020). Comparison of Machine Learning Methods and Conventional Logistic Regressions for Predicting Gestational Diabetes Using Routine Clinical Data: A Retrospective Cohort Study. Journal of Diabetes Research, 2020(4168340). https://www.hindawi.com/journals/jdr/2020/4168340/